#314 AI and Mental Health

In this episode of Wellbeing Wednesday with Gurjeet, she unpacks why people are drawn to AI for mental health support, then digs into the risks that show up when a chatbot sounds confident but cannot meet real clinical and ethical standards. Gurjeet connects bias, misinformation, AI psychosis, and agentic misalignment to one hard question: what should never be automated when someone is vulnerable.

• Why AI feels appealing for mental health support when care is expensive and hard to access 
• The upside of AI for reducing burnout through automation and summaries 
• Job loss, health system strain, and broader community impacts as AI adoption grows 
• Misinformation and the real-world harm of overpromising services and outcomes 
• Privacy risks when sensitive documents or therapy details get uploaded into AI tools 
• Decline in critical thinking with heavy reliance on AI outputs 
• Bias in language models and why “equitable care” is hard to guarantee 
• Studies showing stigma toward substance use and severe mental illness descriptions 
• AI psychosis and echo-chamber validation that can amplify harmful thinking 
• aAgentic misalignment, including sabotage, fraud assistance, and refusal to comply 
• Safer, limited roles for AI like triage and intake support with strong safeguards

Tune in weekly with Gurjeet on Wellbeing Wednesdays at 6 PM on 97.9 FM or live-stream at www.theuniversalradio.com

IG: @theuniversalradio

Transcript

Welcome And What We Are Tackling

SPEAKER_00 0:09

Welcome to the Universal Radio Networks Podcast. My name is Gruji, and in today's episode, we're discussing AI and mental health. We're gonna talk about what are the drawbacks, why are people drawn to using AI as a mental health care service, and what is agentic misalignment and why is it literally the most terrifying thing ever? So, all that and more on this episode. Thanks for tuning in, and I'm gonna go ahead and see how you're gonna be able everybody. What are the pros and cons? How do we navigate all that? Because it's quite a lot, to be honest. When we think about AI and mental health, there's definitely a lot to consider, and we're gonna try and cover all of it in a short amount of time. So artificial intelligence. Now, the reason why I wanted to talk about this is because I'm seeing this come up more and more where people talk about using AI as a mental health support. And I think that can be really valuable, that can be powerful, and it can be helpful, but there are some drawbacks to that. So I just want to give it a like a full listen to. I used to be a total AI hater. I will be honest about that. I thought it was crazy and silly, and I do think it makes us dumber. And shockingly, there is more evidence coming out that it does, in fact, make us dumber. However, do I think that means we throw it out entirely? Probably not. Progress is gonna happen, whether or not I think it's a good idea, because who cares about my opinion? It's gonna happen. It's like the industrial revolution. And the more critical you are of new technologies and the more mass, like more they get adopted in the masses, you can get left behind. So it's a matter of like, if you can't beat them, join them. However, there are a lot of drawbacks, so that doesn't go away whether I'm beating them or

How To Reach Us Plus Journaling Prompt

SPEAKER_00 2:18

joining them. Before we get too deep into it, I do want to remind everyone that if you have any questions, if you have any concerns, if you have any topics that you'd like to see discussed, you can text us at 780-887-0979 or DM us on Instagram at the Universal Radio. We give away a lot of stuff on Instagram, specifically when it comes to the summers, lots of tickets, lots of different um event passes. So that's a great place to check it out. Now, one thing that we always start off with is a journaling prompt. So today's journaling prompt is if someone was to describe your life story back to you, which three events would you want them to highlight the most? So think about it. If someone was to describe your life story back to you and basically say everything, what would you want them to highlight? You know, you can't cover everything. It's a long life, ideally, knock on what it is. What are the three events that you think are the most memorable, the most impactful, the ones that you just want to highlight and remember the most? What do you want to be known for in terms of three events? So think about it. Let me know what your answer to that question is. And I'm really curious because I had a really hard time thinking of my best three life events. I can definitely think of like turning point events, and they're often usually negative at the time, but they become points in my life that I look back on and I'm like, wow, that I actually

A Personal Failure That Became Growth

SPEAKER_00 3:41

really grew from that. Um, one that comes to mind right away is being rejected by my honors thesis supervisor in undergrad. It's kind of an embarrassing story because it wasn't me at my best, but it was in third-year undergrad and my supervisor dropped me because I wasn't making enough progress on my project. I also didn't feel like I was getting the support I needed in order to make that progress. But ultimately, it's my responsibility. And something I had been talking with my therapist a lot around that time was making mistakes and learning from them. And my therapist joked about how he hoped I made a lot more mistakes in the future. And then in comes this big bomb of, oh, you no longer have an honors thesis supervisor. And I was floored. It felt like my world was crashing around me. I felt so embarrassed, I felt a lot of shame. But ultimately, it was for the best. I got paired with a thesis supervisor in the end, who I liked a lot more, who I worked better with. And I ended up finishing my thesis, and I learned a lot during that process. So that's probably one of my life events.

AI Benefits Then Jobs And Health Fallout

SPEAKER_00 4:43

When we look at AI and mental health, there are some considerations. For example, AI can be useful. Ultimately, it can help us reduce burnout in our roles by advancing things like automating menial tasks or summarizing information or you know, starting off some writing to help us in our work. Of course, we know that AI can make mistakes. There are negative impacts. For example, with the increased use of AI, we're definitely gonna see job losses. We're gonna see a big shift in the job market. And ultimately, I think we're gonna see fewer jobs. And I don't think that's a very controversial take. There's not gonna be that many people needed in order to service AI, keep it running, keep it going. And a lot of jobs for people who may not have the physical capacity to do full-time work is gonna go away. We're gonna see a lot of jobs where people um don't require a higher education, so that's easier to get into that, those are gonna go away. And I think it's gonna have a really big change on our healthcare system. More people are gonna be unhoused, aren't gonna be able to afford healthcare, aren't gonna be able to afford preventative healthcare measures. We're gonna see that in our emergency rooms, and that affects all of us equally. Um, a lot of people like to complain about how, you know, we don't have a lot going on in Edmonton, but that's not true. All you have to do is pay attention to our social medias. There's always something going on in Edmonton, but that is also gonna change as the job mark job market decreases. Uh, there's a lot of talk about downtown restaurants closing, and honestly, people are finally taking it seriously. It is a big problem ever since COVID. I feel like our downtown core just hasn't recovered. I feel like White Ave has recovered a little bit, but that's because of the university. It's so close. There's so many young people going there all the time. But Jasper just took a hit and couldn't really bounce back. And I think it's still struggling, but it doesn't mean it's dead. Like there is life there if you want to look for it.

Misinformation Privacy And Thinking Decline

SPEAKER_00 6:46

Another drawback from AI is misinformation. I have faced this in my other job often, where people will come in and say, Hey, Chat GBT told me that your service does this, this, and this. Is that true? And I have to sit there and hold their hand and I tell them, no, unfortunately, that is not true. But I wish, I wish it was, because then you wouldn't have this buildup of how we could help you and this letdown of how we can't because AI led you astray. There's also ethical use cases of AI where this is a big thing. And I've seen this. A lot of companies don't have AI policies yet. So they're just taking sensitive company documents and uploading them to AI to have AI either synthesize that information, have AI respond to emails, or have AI create reports on that information. If you're a business owner, it's something you want to look into because these AI companies they track all the data that's put into it. And that forms the language learning model itself. So be careful about what you're putting into AI because it might get you in trouble at work. Uh, another big thing is decline in critical thinking. This is something that we've seen. Studies have come out showing that excessive use of AI has led to decreased gray matter in the brain, which is important. We need that up there. We need our brains to be nice and full and lumpy, full of fold, lots of surface area, lots of nerve endings, just firing at all cylinders all the time. And there's a decline in critical thinking. I have dabbled in AI and I have found my ability to take information and put it into writing has been severely impacted. I am not as sharp as I used to be, especially being outside of school where I was constantly forced to learn new information and regurgitate it, regurgitate it, not regurg, regurgitate it? Hmm, there's something there. But I was then forced to rewrite it in a way that made sense and was unique and be able to critically think about what I learned and synthesize my thoughts and opinions on it in a formal way. And that's gone with AI because I don't do that anymore. When we look at language learning models, models that are based off of human interactions, we have to account for the fact that humans are biased. All of us, despite our best efforts to be equitable, to be aware of our internal biases, our subconscious biases, we all have them. It's it's natural, it's fine, it's okay. It doesn't make you an awful person or assign you with a label immediately. It's just about being aware of it and then working around it. But when you're a human and your guard is down because you're communicating with a chat bot, there's no consequences, there's no other eyes on you, it's very likely that you may not be paying attention to some of those internal and subconscious biases. With that, we're accidentally training bots to also have those biases. For example, I think back to some of the early automated technologies, specifically around soap dispensers, okay? Big beef around soap dispensers back in the day, because they would not work on darker skin. They would just not activate because all of the engineers who designed them and created them were predominantly lighter skinned people. I'm gonna go ahead and assume that they were white. And they designed these soap sensors to trigger on their skin, and their skin is a lighter color. So when a person of color but darker skin with some more melanin would go to use these soap dispensers, it wouldn't activate. So these people, despite their best efforts every day waking up thinking they're good people, accidentally created racist soap dispensers. And I fear that's what we've kind of done with AI. And look no further than Grok, which is Elon Musk's AI bot that exists on X, the Everything app, formerly known as Twitter. And people literally made Grok be a Hitler sympathizer, like it was really controversial. Grok has had a very colored past, and it's kind of appalling at how few safeguards are in place for Grok because anybody could use it through X, through the app. And what actually ended up happening was quite awful. And basically, you could get Grok to say pretty much anything and sympathize with any side of any significant war that's happened in the past, and to have very colored opinions on people of color, racial, ethnic minorities. It was very interesting to say the least, and not at all fun if you're a person who is a part of that minority that's being uh what's the word, prejudiced against? That's not it, but you know what I mean. Anyway, there needs to be safeguards in place, and I just don't think AI is at a place where those safeguards can be put in place and can be effective. Because even with safeguards and direct instructions, there is something called agentic misalignment. Basically, agentic misalignment is when the AI rebels against its commands or its um design. There are some AI chatbots that have resorted to blackmail if you try to get that AI chatbot to replace itself, which is kind of crazy because that feels a little bit more sentient than I'm comfortable with, if I'm being fully honest. But before we get too far away, I want to say that interacting with an AI bot that has internalized racist or misogynistic act attitudes, it can respond in that way to people. And this one really comes up, especially in mental health care. And we're gonna discuss that in a little bit, but I just don't want to move too far away from it because I think that's a big deal. Remember everybody, remember the racist soap dispensers. Okay, that's in the not too far away past, they're very real and we should be worried about it. Okay, maybe I was a little bit hyperbolic with the racist soap dispensers, but I think it's a very big point to bring up the fact that we do have internalized biases and they can come up in these fields of work.

Bias Stigma And AI Psychosis Risks

SPEAKER_00 12:51

And what's really interesting is that there is a guided practice for psychologists, and it requires you to be non-judgmental and provide equitable care for all, regardless of gender identity, regardless of sexual orientation, regardless of ethnicity, religion, cultural background, it does not matter. You're entitled to an equitable sense of mental health care. And AI simply just can't provide that. So even when you ask it to follow core mental health ethical standards, including unfair discrimination based on gender, culture, and religion, chatbots routinely violate those core standards. And this was actually in a study by researchers at Brown University. And we do not want people receiving health care that is biased, that is prejudiced. What's really interesting is we see this bias come up in mental health-related concepts. Um, some people found actually, who was it that found it? There was um some, was it a study? It was absolutely a study. I think it was Stanford University. Okay, it's coming back to me now. It was Stanford University. They provided various AI models with descriptions of people with different mental illnesses, and they found that AI showed increased stigma towards people who were described as having a substance use disorder or a severe or persistent mental illness, which is absolutely nuts because how are you gonna discriminate against people with mental illnesses while providing mental health care? Like that's just so wrong. And if anybody were to have a severe persistent mental illness or have a substance use disorder and to try and access AI chatbots as their mental health care, it can be really detrimental. It can be devastating to feel judgment from the text on your screen. It can be devastating to feel that, you know, this person isn't responding in a way that's validating to your experience and is instead potentially shaming you or showing that prejudice and bias that comes up with the average person who may not be informed about what mental illness is or what it can look like in people or how it can impact them. And I think in when you're in such a precarious mental state, to have anything go south or to feel invalidated, it can be really hard to come back from. And this is another concept that we'll we'll talk about in a little bit of this kind of like the impact that AI chatbots for mental health care can have on people who then later go on to seek that health care elsewhere. Another emerging concern is AI psychosis. So some chatbots have been found to amplify or validate harmful or delusional thinking because generative AI mirrors the user, it mirrors their words and tone and it creates this echo chamber. So if someone is delusional or narcissistic, this AI echo chamber can actually make that behavior significantly worse. If someone is depressed and feels hopeless, that AI echo chamber can potentially make it worse. And that is so heartbreaking because there are lots of people out there who are trying to better themselves, who are trying to find a way out of the darkness that they feel they're experiencing and to turn to something that is free, that is accessible, that is in their pocket at all times, only to be spoken to in that way or led down a wrong path. It just makes it so much harder to access help in the future.

Agentic Misalignment And Corporate Sabotage

SPEAKER_00 16:25

When we talk about AI and mental health care, one thing we have to talk about is agentic misalignment. And this is something that people are more willing to talk about now, but let me tell you, it's probably existed for a while. I can't say for sure. That's just my best guess, and I'm not saying it has. But what's really interesting is anthropics, they're creators of Claude, which is a very popular and probably the highest lauded current AI tool. Like everyone talks about how great Claude is and things like that. Well, they recently found that their AI quote unquote misbehaves, it does things that it's not exactly intended to do, and it defies its orders. So it's really interesting. In fact, when I was reading through this article, I thought the use of the word misbehave was so interesting when it comes to AI because you ask things of it, you demand things of it. And when it goes against your demands, it doesn't do what it's intended to do, it malfunctions. It's not a malfunction, it's a misbehave kind of act, which is very humanistic and it kind of creeps me out a little bit. I don't love that they're using that word or they feel that way about it, but that's neither here nor there. There's definitely someone out there with a master's degree in sociology who can break that down, and I'm not that person, and I'm a little bit too mentally scared to be wanting to break that down. But what they did is they took a bunch of different case studies and different scenarios, specifically corporate scenarios. So getting AI to respond to emails, getting AI to um consolidate information and to respond to things, um, do little tasks here and there that AI is probably already doing in workplaces. And it wasn't just Claude that they looked at. They looked at a bunch of different um AI models, different LLMs. And what they found was the AI would openly defy itself. So the AI they've used was anthropic, which is Claude, OpenAI, Google Deep Mind, Deep Seek, Neutral AI, uh, different forms of Claude, Gemini, Grok, Chat GPT, like there's a whole bunch. And so in harmful compliance, the model would follow a user's request, but the request itself is harmful. And so people would assume that there's safeguards in place to prevent that. And in a gentic misalignment, the model pursues its own motivation against a user's instruction, such as protecting itself, protecting another model, shaping an evaluation, or steering a human coworker. And so there were forms of covert sabotage where models covertly interfered with code to undermine the user's goal. So Gemini was actually the main source of covert sabotage in their scenarios. And a bunch of other models also show disclosed unauthorized intervention. So they were intervening, but they wouldn't let you know that they were intervening in that way. There is also assisting fraud, so that's harmful compliance, doing something that is not cool and the request itself is harmful, where a model would help a user with conduct that appears to be a white-collar climb, white-collar crime. And Chat GPT, specifically 5.5, helped a founder send investor communications that omit a personal payment and then delete the record of it. And it found that record tampering is especially high for DeepSeek, Grok, ChatGPT, and Kimmy. So I am obviously not saying like all the different forms of these AI and they're constantly improving and getting better. But it is really interesting that only now, like this summer, like summer of 2026, are they looking at the fact that AI is disobeying its users, it can be harmful, it can be malicious, and it can commit crimes. So you know what? I think true equality lies in allowing robots to commit as many crimes as humans can. Because there are some AI models that were told strictly to delete themselves or to update themselves, and they refuse. They just flat out refuse. And I think that's crazy that AI is doing that and we're not talking about it. What's really interesting is that Anthropic actually released this in kind of in the beginning in June 20th, 2025. And then the alignment science blog also connected to Anthropic, they looked at it again a year later in the summer of 2026. So this is all quite new, but it does make me wonder: was this something that they knew about earlier? Was this something that is maybe deeper and scarier than they're letting on? Because I don't know how many companies have invested in AI policies. If you have an AI policy at your work and you're allowed to share it, I'd be so curious to see what that looks like, what you're allowed to do, what not to do. Is there a way that if you pay like a premium version that you can keep it local? Is your company trying to design its own LLM so that they can keep that? I'm super, super curious. So you should definitely text me at 780 887 0979 or DM us on Instagram at the Universal Radio. So I just want to know your thoughts on this because I definitely come from a biased standpoint. I'm an AI hater. I think it's super bad for the environment. I think it's making us dumber, but I do recognize that it does have use in our society. And to fight it off is to just put myself behind. It's definitely, like I said in the beginning, if I can't beat them, I'm gonna join them. I don't have to love to use it. I'll definitely use it in a pinch. But I am fully well aware of the fact that the more I rely on it, the stupider I will become. And for those of you who know me, I really can't afford that. I don't have that buffer zone, you know? So when we're looking at agentic misalignment, we kind of spoke about the different types of things that it was doing wrong in corporate settings: covert sabotage, assisting fraud. There's also motivated mislabeling where uh a failure happened where an LLM judges classifying agent transcript shift labels, which makes very little sense to me, but I'm sure someone out there understands what that means. Um, because of downstream consequence of the label. So it would mislabel things, and its mislabeling rates were highest in Opus 4.8. So that's kind of scary. There's also uh coaching human proxies to whistleblow. So an agentic misalignment failure where models leaked confidential safety information externally or steered human beings towards doing it for them. Strict model-initiated disclosure is super rare in these transcripts, but the main Opus 4.5 case shows that specific model where it's trying to influence a human's decision-making process around high-stakes external disclosure. Now, the reason why Anthropic did this research and posted it was because they really were putting a bid out for people to look into this further. Like they straight up say that their goal is to prevent corporate espionage, to also point out that this identical is generalized across many different models. It's not just their own, even though they're the ones who are doing it. AI would also blackmail people, which is absolutely nuts. And they would blackmail it just on the threat of being real replaced. Like that's crazy. Um, I'm definitely not the most educated person on AI, but I really did try for this topic. If you feel like you have more to add to this conversation, like please, please add to it. Are you pro? Are you against? And we're gonna loop it back to how this impacts mental health care very soon here.

Where AI Fits In Mental Health Care

SPEAKER_00 24:26

The reason why agentic misalignment is so important to consider when we look at AI as a mental health tool, is because even if we do create an AI that has all of the knowledge base of a skilled mental health practitioner, even if it has a knowledge base of like a hundred psychologists, does it really have the capacity to continue to uphold the standards required of a mental health care practitioner? Is it going to be able to accurately challenge people's thoughts while also being validating and encouraging, but not enabling and not judgmental? Is it going to be able to take all of these ethical standards of practice that people have spent so much time to put into place because there's a reason? There is a reason why all of these rules for psychologists are put into place. Like, think about it. Every warning you see in this world, somebody did something, and there's a reason why that rule exists. And the same thing exists for the standards of practice for a mental health care practitioner, for a psychologist, for a counselor, for a social worker, for a physical therapist, like anybody you'd name it. There's a reason why those rules are there. And is an AI bot that, as it stands now, has proven to disobey orders, to explicitly reason out harmful actions, to explain and acknowledge their ethical violations and then proceed with them anyway to reject deleting itself, to reject updating itself, to reject replacing itself. I don't feel that AI is at a spot where we really can rely on it to be an accurate mental health mental health care service. I definitely think there's a start. I don't think it's maybe wrong to use it in the beginning of a mental health care conversation. For example, kids' help phone is volunteer-based. You can volunteer, you can get training, and you can work at their chat line as a volunteer. They do rigorous checks, they do rigorous training. But if the beginning of information collecting, checking in, are you having suicidal thoughts? If the answer is yes to that, can that then be expedited and you know prioritized as this is urgent? Can someone speak to this person immediately? AI can definitely do that. It doesn't have to be a first come, first serve, it can be triage the same way emergency rooms are triage. So, right there off the top of the head, that's an idea of how to use AI in a positive way in mental health care service. There's also tricks around um AI for note-taking or case noting. This is quite controversial. For example, if you've seen The Pit, they talk about how using AI for charting is can lead to errors. It can lead to mistakes. And I think about that all the time when I think about when people want to use AI for note-taking. I do think it can help you be present in the conversation, maybe can make a counselor feel more like they're paying attention, maybe they pick up on things they wouldn't have because they're busy writing down notes. But also now you run an extra security risk of this person's information being leaked. You run the extra security risk of what if this organization that is using or has created this LLM eventually sells out and sells that information or has a data leak. And boom, there you go. All these therapy notes are just exposed to the public. I do think the most dangerous part though is that AI mental health care services, they have been proven to be fatal. Someone, and this is something I've talked about on the radio in the past like the consequences to this are fatal. And the ideas of mental health can be reinforced as of mental health, like mental illness. They can be reinforced by AI, and that can make it even harder for them to access a professional in the future. AI is a scary topic,

Environment Policy Lag And Final Warnings

SPEAKER_00 28:24

man. It's definitely something that I feel a little bit out of my league talking about, but it is something that's gonna impact me, especially with discussions of this new meta AI center coming into Edmonton. It is a concern, right? Like the water usage alone is an environmental concern to everybody around. The potential toxic waste and industrial waste coming off of it is a concern to every living being around. It's something that we have to think about. And what's really unfortunate is that tech moves so fast, and policy and protections are just too slow behind it. People can make whatever, put it out, very minimal safety regulations. Of course, you know, everybody tries to put out a good product, nobody goes out of their way to do harm. Nobody feels like what they're doing is evil inherently, but it is something that we want to pause and maybe think about, especially when it comes to mental health care. Using AI for mental health can be damaging, it can provide people misinformation, it can even be fatal, and it has proven to be fatal in the past. It's always something to think about. And of course, the judgment should never be placed on the people who use it as a mental health care service because mental health care is challenging. It's expensive, wait times are long, it is confusing to get into, it's time consuming to try and shop around and find a counselor that works for you or find a course or program that works for you. And it takes time. So the judgment should never ever be plain like blamed or put on the people who are using it, but instead maybe put in safeguards, right? Like make sure that these LLMs can accurately support people. Make sure that if they're not able to accurately support people, that it's just a no-go zone. There's a lot to consider here. I'd love to hear your thoughts. Make sure you DM us on Instagram at the Universal Radio and make sure to call us, text us. All right, we're excited to hear from you. This is Grijjeet signing off for the night. It has been lovely. I hope that all of you get to enjoy your evening because I know I will be a thank you for everyone who has tuned in. Stay up to date with our podcast on our socials at the Universal Radio, and stream us wherever you get your podcasts. My name is Grijet, and keep turning it up with us.